AI Detectors Can't Read Chinese or Japanese: A 62-Text Test
We ran 62 Chinese and Japanese texts, half of them AI-written, through ZeroGPT's API. Every text it accepted scored 0% AI, because it never split the text into words.
Short answer: not the ones we tested. We ran 62 Chinese and Japanese texts through ZeroGPT's detection API, the same check our own AI Detector uses. Every text it accepted scored 0% AI, including 30 written by AI. The score did not mean "human-written". The detector never split the text into words, so it had nothing to analyze.
This matters if you write in Chinese or Japanese, or teach students who do. A 0% from these detectors is not a clean result. It is no result.
How we tested
- Human-written Japanese (25 texts): 12 passages from Japanese Wikipedia articles, taken from revisions saved before January 1, 2020, and 13 public-domain essays by Torahiko Terada and Ukichirō Nakaya from Aozora Bunko. Both sources predate today's chatbots.
- AI-written Japanese (25 texts): 13 by Doubao (seed-2.1-pro) and 12 by DeepSeek, on student-style tasks: university reports, short essays, book reports, explanations, opinion pieces, self-introductions, business emails, blog posts and a seminar talk.
- Human-written Chinese (6 texts): passages from Chinese Wikipedia revisions saved before January 1, 2020.
- AI-written Chinese (6 texts): by Doubao: a course paper, a science explainer, reading notes, an argumentative essay, a work email and a blog post.
- Length: 378 to 1,152 characters per text, roughly the length of a short assignment.
- Detectors: ZeroGPT's detection API, one call per text, with no editing. We then sent all 62 texts to a second service, Ryne's AI score API.
- Control: one English paragraph written by AI.
All runs took place on October 28, 2026. We did not publish the source texts; the method above is enough to repeat the test.
Results
| Group | Texts | Accepted by ZeroGPT | ZeroGPT AI score | Ryne AI score |
|---|---|---|---|---|
| Japanese, human-written | 25 | 25 | 0% for all | 0% for all |
| Japanese, AI-written | 25 | 23 | 0% for all | 0% for all |
| Chinese, human-written | 6 | 6 | 0% for all | 0% for all |
| Chinese, AI-written | 6 | 5 | 0% for all | 0% for all |
| English, AI-written (control) | 1 | 1 | — | 100% |
ZeroGPT rejected three texts outright: two Japanese blog posts and one Chinese science explainer. It scored every other text 0%. Ryne accepted all 62 and also scored every one 0%, while it scored the English control 100%.
So on these texts, neither detector caught a single AI-written passage. It also "cleared" all 31 human passages, but only because it cleared everything.
Why every score was 0%
ZeroGPT's API reports how many words it found and which sentences it flagged. We recorded these details for all 11 Chinese texts it accepted and for 4 of the Japanese texts (two human, two AI):
- Words counted: 0 to 23 for the Chinese texts and 0 to 2 for the Japanese texts. Each text had several hundred characters.
- Sentences found: 0 in every one of them.
- Verdict text: often "Your Text is Most Likely AI/GPT generated", printed next to a score of 0%.
For comparison, Ryne, which reports the same fields, read the English control as 117 words and 7 sentences and flagged all of them.
Chinese and Japanese do not put spaces between words. A detector that counts words by spaces sees a paragraph as one or two very long "words". ZeroGPT's score is the share of AI-written words in the text, so with no words to judge, the score falls to 0%. The verdict text and the score then contradict each other, which is a sign the result is not meaningful.
Ryne returned the same fields and the same verdict wording, so it behaved the same way.
What the vendors say
- ZeroGPT says it aims to support all languages (ZeroGPT API page). Its detection API takes only the text, with no language setting (API reference).
- GPTZero lists Chinese and Japanese among its supported languages (GPTZero help).
- Copyleaks lists Japanese (
ja) and Chinese among the languages its AI detector supports (Copyleaks docs).
We tested only ZeroGPT and Ryne. We have not tested GPTZero, Copyleaks or Turnitin on Chinese or Japanese text, so we make no claim about them.
What this means for you
- If you write in Chinese or Japanese: a 0% from a detector that counts words by spaces does not show your text reads as human. Check whether the report found any words or sentences before you trust it.
- If you teach: do not treat a 0% on Chinese or Japanese work as evidence either way. Ask which detector was used and whether it supports the language.
- If you check English text: this test does not change our English results. On 60 English texts, ZeroGPT flagged 58 of 60 unedited AI drafts (our ZeroGPT test).
Our own AI Detector and ChatGPT detector use ZeroGPT, so this applies to them too. Use them for English and other space-separated languages, not for Chinese, Japanese or Korean.
Limits of this test
- Two detectors. Other detectors may handle Chinese and Japanese properly. We did not test them.
- Small sample. 62 texts, one run each. The pattern was the same in every group, but the sample is small.
- No Korean. Korean uses spaces between words, so it may behave differently. We did not test it.
- One date. Detectors change. These results describe the APIs as they behaved on October 28, 2026.
FAQ
Can ZeroGPT detect Chinese AI text? Not in our test. It scored all 11 Chinese texts it accepted at 0%, including 5 written by AI, and counted at most 23 words in each.
Can ZeroGPT detect Japanese AI text? Not in our test. It scored all 48 Japanese texts it accepted at 0%, including 23 written by AI. In the four we inspected in detail, it counted at most 2 words.
Does a 0% AI score mean my text is human-written? Not on Chinese or Japanese text from these detectors. Check whether the report counted words and sentences. If it found none, the text was not analyzed.
Which AI detectors support Chinese and Japanese? GPTZero and Copyleaks say they do. We have not tested them on these languages yet, so we can't confirm how well they work.
*Test run on October 28, 2026. Vendor statements checked the same day.*
